Anda di halaman 1dari 18

A method to assess demand growth vulnerability of travel times

on road network links


David Watling

, N.C. Balijepalli
Institute for Transport Studies, University of Leeds, Leeds, England, United Kingdom
a r t i c l e i n f o
Keywords:
Reliability
Networks
Travel time variability
Vulnerability
Demand growth
Day-to-day variation
a b s t r a c t
Many national governments around the world have turned their recent focus to monitoring
the actual reliability of their road networks. In parallel there have been major research
efforts aimed at developing modelling approaches for predicting the potential vulnerability
of such networks, and in forecasting the future impact of any mitigating actions. In prac-
ticewhether monitoring the past or planning for the futurea confounding factor may
arise, namely the potential for systematic growth in demand over a period of years. As this
growth occurs the networks will operate in a regime closer to capacity, in which they are
more sensitive to any variation in ow or capacity. Such growth will be partially an expla-
nation for trends observed in historic data, and it will have an impact in forecasting too,
where we can interpret this as implying that the networks are vulnerable to demand
growth. This fact is not reected in current vulnerability methods which focus almost
exclusively on vulnerability to loss in capacity. In the paper, a simple, moment-based
method is developed to separate out this effect of demand growth on the distribution of
travel times on a network link, the aim being to develop a simple, tractable, analytic
method for medium-term planning applications. Thus the impact of demand growth on
the mean, variance and skewness in travel times may be isolated. For given critical changes
in these summary measures, we are thus able to identify what (location-specic) level of
demand growth would cause these critical values to be exceeded, and this level is referred
to as Demand Growth Reliability Vulnerability (DGRV). Computing the DGRV index for
each link of a network also allows the planner to identify the most vulnerable locations,
in terms of their ability to accommodate growth in demand. Numerical examples are used
to illustrate the principles and computation of the DGRV measure.
2012 Elsevier Ltd. All rights reserved.
1. Introduction
In the last decade, two complementary considerations have emerged in the analysis of road transport networks:
Monitoring reliability Empirically, typically based on measurements of travel times, how reliable has a transport network
been in the past in delivering repeatable levels-of-service to its users? Such a consideration has par-
ticularly attracted the attention of local and national government agencies, who wish to assess
changes over time (days, months, years) to such level-of-service repeatability, and to gauge the
impact of any past measures that have been taken.
0965-8564/$ - see front matter 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.tra.2012.02.009

Corresponding author.
E-mail address: D.P.Watling@its.leeds.ac.uk (D. Watling).
Transportation Research Part A 46 (2012) 772789
Contents lists available at SciVerse ScienceDirect
Transportation Research Part A
j our nal homepage: www. el sevi er . com/ l ocat e/ t r a
Forecasting reliability How reliable is a transport network likely to be in the future, taking into account its connectivity
(availability of alternatives), potential failures (e.g. in capacity), and any proposed remedial mea-
sures to improve its reliability? This has largely attracted the attention of the research community
in developing suitable network modelling methods, but clearly holds signicant interest for trans-
port planners who wish to select among potential measures and policies based on their potential for
reliability improvements.
If our focus is relatively short-term, then in monitoring a network we may be interested to know in detail what has hap-
pened to reliability over the past few hours, days or weeks, and our forecasts may pertain to only the immediate future. In
the longer-term, on the other hand, we may be interested in monitoring reliability more widely in order to benchmark it
against the levels seen in previous years, or to forecast how it might change over future horizons in the order of years. While
the practical division between short-term and longer term is arbitrary, a key element that distinguishes them is that in the
longer term we have the possibility of systematic, confounding changes to the underlying transport system, which may have
major consequences for any comparisons made between states of reliability observed or predicted many years apart from
one another. This kind of consideration serves as the context in which the present paper is written, with a focus beyond
the short-term, and motivated by both the monitoring and forecasting issues described above.
Looking to the literature, it is apparent that this emerging research eld of transport network reliability has provided a
new focus for transport planning (e.g. Bell and Cassir, 2000). Rather than only focusing on typical or average performance
of a transport system, in studies of network reliability we are able to focus on the impact of variability in the factors that
affect the system, be they frequently-occurring or abnormal events. While the origins of the eld drew from the study of
major natural disasters such as earthquakes, the remit has broadened to encompass a range of factors affecting both the
capacity of the transport system (from accidents and man-made disasters to weather conditions), the demand for travel
on the system (from normal daily uctuations in commuting patterns to special events), and the behaviour of the actors in-
volved in the system (from individual travellers adapting their behaviour to agencies providing information, planning reme-
dial measures or designing robust transport networks). For recent reviews the interested reader is referred, for example, to
the literature cited in Clark and Watling (2005) and Sumalee and Kurauchi (2006).
Within this eld, we may identify two complimentary perspectives that are developing. The rst of these perspectives
begins by representing the variability in the system components, typically through some probability distribution, and aims
to compute some measure of reliability that the system would attain. For example, in studying Travel Time Reliability we
may assume link capacities or origindestination demands to be stochastic, then based on some behavioural response of
travellers to such stochasticity we may infer a probability distribution for the travel times on the network, and thereby ob-
tain a measure such as the probability that the OD travel time will exceed some value (for examples, see Clark and Watling,
2005; Shao et al., 2006; Sumalee and Watling, 2008). Many variants on this approach exist, ranging from the initial work on
Connectivity Reliability, where our interest is in the probability of a given OD pair remaining connected even when some
links are unavailable in a degraded network (Bell and Iida, 1997), to the work on Capacity Reliability by Yang et al.
(2000) and Chen et al. (2002) whereby the probability of catering for a given level of demand is computed in the face of sto-
chastic network capacities. In spite of the many variations on this theme, the key common element is the premise that we
have sufcient information to be able to specify both (a) the component probability distributions and (b) the responses of
travellers to the variable conditions.
The second perspective starts from an alternative viewpoint that, instead of trying to estimate how unreliable is the sys-
tem under assumptions about the unreliability of its components, the focus should be on identifying potential weak-points;
this is the area of Vulnerability Analysis. Such works aim to identify vulnerable elements of a network based on what could
happen under pessimistic assumptions (e.g. Bell, 2000; DEste and Taylor, 2001; Berdica, 2002; Nicholson and Dalziell, 2003;
Taylor et al., 2006; Jenelius et al., 2006; Szeto et al., 2006; Jenelius, 2009). As in the rst theme earlier, this grouping of papers
comprises quite a diverse range of approaches, though commonly focused on the problem from a supply (capacity) point of
view. For example, DEste and Taylor (2001) and Taylor et al. (2006) analysed network-related socio-economic impacts based
on reduced accessibility in a degraded network, whereas Nicholson and Dalziell (2003) considered the impacts of network
degradability in the event of natural disasters such as volcanic eruptions. Berdica (2002) considered vulnerability from the
viewpoint of how well the transport system performs when exposed to interruptions to critical links, general capacity reduc-
tions and variability of the demand, using equilibrium models to assess the sensitivity of networks to such potential inter-
ruptions. Bell (2000) and Szeto et al. (2006), on the other hand, directly represented an evil entity that aims to degrade
system performance, suggesting that we might evaluate worst-case scenarios by examining a game played about between
the evil entity and the travellers in the network. The common theme in these papers is that we do not aim to evaluate how
reliable is the transport system, and therefore do not need to provide as input the probability distributions of the degraded
components. This is particularly attractive in cases of rare events, where it is extremely difcult as a modeller to give a reli-
able estimate of the input probability distribution.
In spite of the differing perspectives of reliability and vulnerability analyses, a common theme running through these
works is the examination of a transport system under conditions of variability or potential failure, but in a context where
the transport system essentially stays constant. That is to say, the laws governing variability or failure (and their impacts)
are constant, be they probability laws, behavioural responses or the background network/demands against which potential
failures are explored. In practice, however, there are likely to be underlying systematic changes to a transport system over
D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789 773
the longer term, and these changes may not be neutral to an assessment of reliability/vulnerability. As we shall discuss later
(Section 2), this has particular practical relevance for organisations wishing to benchmark against previous years perfor-
mance, or to plan for robust systems in the longer term.
The focus of the present paper is from the perspective of vulnerability, yet we shall make several departures in our meth-
odology from approaches proposed previously, especially motivated by the issue of longer-term systematic change men-
tioned above:
v Firstly, our focus will be on the role of demand in network vulnerability, whereas previous vulnerability analyses have
almost exclusively focussed on link capacities.
v Secondly, we shall explore the impact of systematic growth in demand (that might occur over a period of years) on vul-
nerability. To our knowledge, this is not an issue that has previously been addressed in either reliability or vulnerability
analyses. In essence, the thesis is that demand growth will lead our transportation system to operate in a regime where it
is more vulnerable to uctuations (in either ows or capacities). Therefore, identifying the links of a network that are
most vulnerable to demand growth should allow us to direct policy interventions to design more robust networks. For
example, we may choose to divert trafc away from an area that has little capacity to accommodate demand growth, even
if conventional economic analysis of current conditions directs us otherwise.
v Thirdly, our assessment of vulnerability will be a stochastic one, in the sense that we shall assess vulnerability not just
with respect to mean performance, but also with respect to higher order moments. In this respect, it might be said that
we aim to assess Vulnerability of Reliability.
v Fourthly, our focus will be on a component-wise, link-by-link assessment of network vulnerability. An alternative
approach, more in line with previous studies, would have been to integrate this link-by-link assessment with a network
assessment through use of some network equilibrium model, e.g. a kind of vulnerability counterpart of the approach in
Clark and Watling (2005). The reason for not adopting this approach is that we wished to devise a method that made min-
imal assumptions. As noted earlier, techniques such as Travel Time Reliability, Connectivity or Capacity Reliability all
need to make two kinds of assumptions: not only assumptions about the probability distributions of the input compo-
nents (capacities, demands) but also assumptions about how the system/travellers respond (typically assuming self-orga-
nisation through a network equilibrium approach). Past vulnerability analyses have typically only removed the need for
the rst kind of assumption.
In essence the thinking behind the approach is that it will be a simple means of providing transport planners with a way
to assess the vulnerability of Travel Time Reliability on a link to growth in demand for using that link. It is anticipated that
this could be the rst stage in an assessment, which may subsequently lead to a network assessment of vulnerability.
The structure of the paper is as follows. In Section 2, we briey discuss the practical context of our work, introducing the
measures of reliability of travel time as used in practice in various countries such as the USA, the Netherlands and England.
We then develop our analytical approach in two stages. Our focus initially will apparently not be on vulnerability at all, but
rather on a retrospective analysis of past, observed conditions. This is the calibration-phase of the model, where we use his-
torical data to estimate the parameters of the model. Thus Sections 27 all take this retrospective, historical perspective. It is
not until Section 8 that we then use the same techniques in a forward-looking context; effectively the methods are the same,
but are instead used to estimate the (future) vulnerability to demand growth of the network links. Preceding this, then, in
Section 3 a theoretical model is proposed to explain ow-related variations in travel times and the role of systematic trafc
growth, alongside residual, unexplained sources of travel time variation. It is then shown, in Sections 46, how this method
may be applied to retrospectively adjust observed travel times for growth in trafc. The essential simplicity of the method,
which can be lost in the algebraic manipulations required for the most general case, is illustrated by reference to a special
case in Section 7, where travel times follow a second order relationship in ows, and where very simple explicit relationships
can be derived. Finally, Section 8 moves onto the nal objective of assessing vulnerability to demand growth, including a
numerical illustration.
2. Context: measures of Travel Time Reliability in practice
In this section, we shall briey set out the practical motivation for our work, by considering the measures and approaches
to reliability presently proposed by several governments around the world, and in particular will examine the extent to
which a longer-term temporal dimension plays a part in such measures and policy objectives.
In the USA, the FHWA (2005) suggests several statistical measures to be used in the monitoring of Travel Time Reliability,
including both simple statistical summary measures such as the standard deviation, coefcient of variation, or 95th percen-
tile, as well as measures in which a traveller is assumed to compare their experience against an expectation, such as the buf-
fer index and planning time index, both of which consider the time cushion needed by travellers to ensure on-time arrival. In
making the case for such measures, Fig. 2 in FHWA (2005) (the gure entitled Reliability measures capture the benets of
trafc management) illustrates how the tracking of such reliability measures over a period of two years, before and after
some network improvement, allows such reliability benets to be estimated. However, implicit in the comparison made
is that the changes are only attributable to the measure introduced; the comparison would not be appropriate, for example,
774 D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789
if there had been signicant systematic change in travel demand levels over the period monitored. This is a perfect example
of the issue of interest to the present paper, namely that simple monitoring of reliability may not be sufcient to isolate the
effects of scheme impacts, if other confounding factors are at work at the same time.
The Rijkswaterstaat (2005) in the Netherlands interestingly start with a policy target that has an explicit time dimension
with respect to reliability, namely a specic ambition to have 95% of rush-hour movements arriving on time by 2020. Given
such a relatively long-term (15 year) target, there is clearly considerable potential for systematic changes in several factors
to confound any monitoring exercise of journey time repeatability. The measure suggested for monitoring reliability over
this period, simply the standard deviation of travel times, will clearly be sensitive to systematic changes in demand that
may make the network more or less sensitive to day-to-day and within-day uctuations in demands and capacities. While
it may be that policies aimed at reducing the overall level of demand on the road network could indeed be said to fall within
the remit of a measure to achieve this ambition, and so be a quite legitimate factor to capture in the monitoring, it is not clear
that any counter-effect of other exogenous factors which may increase demand are also intended to be part of this
assessment.
In England, the Highways Agency is responsible for monitoring and delivering an agreed level of service on the strategic
road network in England comprising Motorways, A roads and trunk roads. The level of service provided is measured by a
certain PSA (Public Service Agreement) target (known as PSA target1) for the strategic network. In 2004, then looking to
the future, this target was dened as effectively that the slowest 10% journeys in the reference year 2007/8 should not be
worse off when compared to the slowest 10% journeys in the base year 2004/5 (DfT, 2007). PSA target1 requires computing
what is referred to as Average Vehicle Delay (AVD) in both the reference year and the base year. AVD is measured in min-
utes per 10 vehicle-miles travelled, and is dened as the sum of the total delay on the slowest 10% of the daytime journeys
divided by the sum of the vehicle-miles, on each of the identied routes on the strategic network over each 15-min departure
period between 6amand 8 pm for each day of the week (DfT, 2007).In order to account for the day-to-day variability in travel
time, the slowest 10% journeys are identied based on a distribution of the average journey times for each route within each
15-min window on each week day. For every route in each route direction, for a given week day within each 15-min depar-
ture period, the total delay is computed as the product of the average ow and the difference between the experienced jour-
ney time in the reference year and the journey time in the base year (August 2004July 2005). Vehicle-miles is dened as the
product of the average ow (for which the travel time exceeds the 90th percentile travel time) and the route length in miles.
Experienced journey time is the average journey time in each 15-min departure period, whereas the journey time in the base
year is based on the weekday off-peak median speed. Reference journey times are set at lower than the speed limits because
on an average, vehicle speeds are less than the speed limits. Finally, average vehicle delay is obtained by dividing the total
delay by the vehicle-miles. The reliability index on which PSA target1 for England is based holds special interest for the pres-
ent paper, since it includes within it an explicit baseline timeframe, of the reliability measured some several years
previously.
In summary, fromthese three examples we see diverse reasons as to why a consideration of systematic change in demand
over time is relevant in a reliability context: (i) identifying from a data monitoring exercise the impact of a past intervention
on reliability (as in the FHWA example); (ii) allowing for future demand changes as part of a longer term policy objective to
improve reliability (as in the Netherlands example); and (iii) as part of a more complex measure of reliability itself which
incorporates an idea of a baseline level of reliability in some past year (as in the example of the Highways Agency in Eng-
land). The purpose of our research, having identied this gap, is to address it by developing a practically useable methodol-
ogy based on a sound mathematical footing, as we shall describe in the subsequent sections.
3. Model assumptions
In monitoring the performance of a highway facility over a course of years, it is natural to examine observations of travel
times, recorded at similar times of day for similar types of day, compared between some base year and a current year. Since
travel times, even when distinguished by time-of-day and day-type, are rather variable from day-to-day, then this compar-
ison naturally takes the formof comparing elements of the distribution of travel times, e.g. the mean, variance and percentiles
between the base- and current-year. In observing any differences in such distributions, a natural objective is to attempt to
attribute them to some underlying factors, and this is the topic of the present paper. In particular, the primary aim of the
paper is to develop a methodology to adjust for the confounding effect of systematic trafc growth over time, as is observed
in many developed and developing countries around the world.
We consider a particular stretch of road, a link, where it is supposed that repeated observations of travel times are made
over a number of days. These observations are grouped according to time periods within the day, and day-types, the latter to
account for systematic patterns within the week, e.g. differences of Saturdays to Mondays. We shall henceforth focus on one
particular time period within the day (representing, say, a period of 515 min duration), and a particular day type, and sup-
pose that the observations for each particular day are used to form a mean for that day for that particular period of the day
(intra-period mean). For example, if a particular day-type were Friday, and we observed all Fridays for a year between the
times of 8:008:15, then we would obtain 52 observations (one for each Friday in the year) of the intra-period mean travel
time for 8:008:15. When we refer to travel time observations below, then, we are referring to the observations of such
intra-period mean travel times over a sequence of days of the same type, for the same period of the day.
D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789 775
The day-to-day variability in these intra-period mean travel times is modelled by a random variable T with two compo-
nents of variation:
1. Flow-related sources of day-to-day travel time variability t(F), where F is a random variable denoting day-to-day variabil-
ity in ow for the period of the day being modelled, and t() is some deterministic (generally, non-linear) function.
2. Residual sources of travel time variability:
T[(F = f ) = t(f ) e (1)
where e is a random variable denoting the additional day-to-day variability in travel time that arises, even given knowledge
of a given days ow level f.
Based on this basic model, several specic assumptions are now made for the purposes of subsequent analysis.
Assumption 1. The function t(), known as the travel time function, is of a polynomial form:
t(f ) = b
0

n
i=1
b
i
f
i
(2)
for some constants b
0
, b
1
, . . . , b
n
.
Assumption 2. The ow f denotes total ow measured in Passenger Car equivalent Units (PCUs) per hour:
f =

K
k=1
k
k
g
k
(3)
where g
k
denotes the ow of vehicle type k and k
k
denotes the passenger car equivalent of vehicle type k.
Assumption 3. Day-to-day variation in ow may be approximated by a Normal distribution:
F - Nor(l; r
2
) (4)
where l and r respectively denote the mean and standard deviation in trafc demand levels.
Assumption 4. eis a zero-mean
1
random variable.
The assumptions above apply to a particular link, for a particular time period of the day; any of the parameters in the
assumptions may vary by link and time period of the day, given sufcient data.
4. Statistical foundation: travel time moments
The method that will be developed for understanding the role of trafc growth will be based on two stages. Firstly we aim
to understand how the theoretical travel time distribution, arising from the assumptions in Section 3, depends on various
parameters of the model (Sections 4 and 5). We do not specically mention a base year and reference year, but it is through
these latter parameters that we will be able to reect changes between the observation years, e.g. trafc growth. This leads to
the second stage, in Section 6, where we are able to compare changes in the theoretical distribution with changes in the ob-
served data between the reference and current year, and thereby derive a method for adjusting the current-year observa-
tions for trafc growth. In the present section, we focus on the rst part of the rst stage of this process, monitoring
changes to the theoretical distributions.
In practice, we can specify a probability distribution in many ways, not just as is standard through the probability density
function; one such specication is through its moments, which has an advantage of allowing a more parsimonious descrip-
tion of given families of distribution. For example, the Normal distribution is fully specied by its rst two moments, the
mean and variance, and for arbitrary distributions we can t the parameters of the distribution to observed moments (by
the method of moments), provided that we have as many moments as parameters. For our subsequent analysis, we shall
propose to make use of a three parameter, asymmetric distribution, namely the lognormal, for representing the distribution
of travel times, and this motivates the later parts of the discussion below to focus attention on the rst three moments of the
travel time distribution (as the tted distribution has three parameters). However, the methods presented can be extended
to any other desired distribution with any number of parameters, by using the general formulae below to generate as many
moments as there are parameters.
1
The assumption that e has zero mean is not restrictive, since if the additional sources of variation (unrelated to ow) have an impact on mean travel time,
then this may be accommodated in the constant term b
0
of t(f).
776 D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789
Now, moments of the travel time distribution about the mean (/
m
; m = 2, 3, . . .) can be obtained from the binomial expan-
sion of its denition:
/
m
= E[(T E[T[)
m
[ =

m
j=0
(1)
mj
m!
j!(m j)!
E[T
j
[(E[T[)
mj
(m = 2; 3; . . .): (5)
The right-hand side of (5) therefore requires knowledge of the travel time moments about the origin:
/
/
m
= E[T
m
[ (m = 1; 2; . . .): (6)
Below, it is important to keep track of these two kinds of moments, those about the mean and those about the origin, as
we shall move between them several times. Note that from (6), the mean is /
/
1
and from (5) the variance is /
2
, and for higher
values of m these moments describe other features of a distribution such as its skewness. Using (6), and taking the conven-
tion that /
/
0
= 1, then we can write (5) as:
/
m
=

m
j=0
(1)
mj
m!
j!(m j)!
/
/
j
/
/
1
_ _
mj
(m = 2; 3; . . .): (7)
Now, let us turn to the specication of the distribution of travel time T under the assumptions specied in Section 3, com-
prising a component related to ow and a residual component. Using a standard statistical identity and by taking expecta-
tions of powers of T[F = t(F) + e with respect to e and then F, we obtain:
/
/
m
= E[T
m
[ = E[E[T
m
[F[[ = E[E[(t(F) e)
m
[F[[: (8)
It should be noted that, while the notation in (8) is standard, it hides a number of subtle points, in particular the meaning
of the expectation operator, where the meaning is taken to be implicit from the context. Thus in writing E[T
m
] the expecta-
tion is clearly the mean of travel times raised to the mth power, taken across the distribution of the random variable T. How-
ever, E[T
m
[F] might seem less obvious in meaning, since it involves two random variables; however, since we condition on F
then this variation cannot be the subject of the expectation, and so the expectation must represent the mean across the con-
ditional distribution of T given F. The result of this expectation is itself a random variable, depending on the distribution of F,
and therefore the outer expectation of this term must implicitly refer to the mean across the distribution of F.
Now, with a further expansion inside the last expectation in (8) we have:
/
/
m
= E E

m
k=0
m!
k!(m k)!
(t(F))
k
e
mk

F
_ _ _ _
=

m
k=0
m!
k!(m k)!
E[(t(F))
k
E[e
mk
[F[[: (9)
For example, let us assume that e[F - Nor(0, w(F)) for some function w(), with this latter function intended to reect po-
tential heteroscedasticity. Since, as mentioned in the opening paragraph, our interest will be in travel time moments up to
order m = 3, then we immediately know from standard properties of the Normal distribution that:
E[e[F[ = E[e
3
[F[ = 0 E[e
2
[F[ = var(e[F) (E[e[F[)
2
= w(F) (10)
leading to some considerable simplication in (9). If we followthis assumption through to compute the rst three travel time
moments about the mean fromthe combination of expressions (9) and (10) above, then rstly for the mean travel time /
/
1
we
have:
/
/
1
= E[e[F[ E[t(F)[ = E[t(F)[: (11)
For the second moment about the mean (variance) /
2
we rst have from (7):
/
2
= /
/
1
_ _
2
2 /
/
1
_ _
2
/
/
2
= /
/
2
/
/
1
_ _
2
(12)
and in parallel from (9) and (10) we have:
/
/
2
= E[E[e
2
[F[[ 2E[t(F)E[e[F[[ E[(t(F))
2
[ = E[w(F)[ E[(t(F))
2
[: (13)
Hence, combining (12) and (13):
/
2
= E[w(F)[ E[(t(F))
2
[ /
/
1
_ _
2
: (14)
Finally, for the third moment about the mean /
3
, we rst have from (7), after some simplication:
/
3
= /
/
3
3/
/
2
/
/
1
2 /
/
1
_ _
3
: (15)
The only term not yet computed in (15), namely /
/
3
, is then given by (from (9) and (10)):
/
/
3
= E[E[e
3
[F[[ 3E[t(F)E[e
2
[F[[ 3E[(t(F))
2
E[e[F[[ E[(t(F))
3
[ = 3E[t(F)w(F)[ E[(t(F))
3
[ (16)
and substituting (16) into (15):
/
3
= 3E[t(F)w(F)[ E[(t(F))
3
[ 3/
/
2
/
/
1
2 /
/
1
_ _
3
(17)
D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789 777
or writing /
2
/
/
1
_ _
2
in place of /
/
2
(from (12)):
/
3
= 3E[t(F)w(F)[ E[(t(F))
3
[ 3 /
2
/
/
1
_ _
2
_ _
/
/
1
2 /
/
1
_ _
3
= 3E[t(F)w(F)[ E[(t(F))
3
[ 3/
/
1
/
2
/
/
1
_ _
3
: (18)
Thus we can collect together the results above into three sequentially dened expressions:
/
/
1
= E[t(F)[; (19)
/
2
= E[w(F)[ E[(t(F))
2
[ /
/
1
_ _
2
; (20)
/
3
= 3E[t(F)w(F)[ E[(t(F))
3
[ 3/
/
1
/
2
/
/
1
_ _
3
: (21)
While w(F) reects the possible presence of heteroscedasticity, we may also in cases accept a simplifying assumption of a
homoscedastic error variance; in this special case, if we let
w(F) = b
2
(22)
for some constant b > 0, then the moment expressions (19)(21) reduce to:
/
/
1
= E[t(F)[; (23)
/
2
= b
2
E[(t(F))
2
[ /
/
1
_ _
2
;
/
3
= 3b
2
E[t(F)[ E[(t(F))
3
[ 3/
/
1
/
2
/
/
1
_ _
3
: (24)
This expression for /
3
may be further simplied, by substituting for E[t(F)] from (23) and substituting for /
2
from (24),
yielding:
/
3
= 3b
2
/
/
1
E[(t(F))
3
[ 3/
/
1
(b
2
E[(t(F))
2
[ /
/
1
_ _
2
) /
1
( )
3
which simplies to a form that shows that in this case the effect of b on the third moment is eliminated:
/
3
= E[(t(F))
3
[ 3/
/
1
E[(t(F))
2
[ 2 /
/
1
_ _
3
: (25)
Now, as we have assumed t() to be of polynomial form, all the expectations of powers of t(F) in (23)(25) above, when
expanded, can be expressed as expectations of polynomials in F. Since:
E[(t(F))[ = E b
0

n
i=1
b
i
F
i
_ _
= b
0

n
i=1
b
i
E[F
i
[ (26)
and for m = 2, 3:
E[(t(F))
m
[ = E b
0

n
i=1
b
i
F
i
_ _
m
_ _
= E

b
m0

mn
i=1

b
mi
F
i
_ _
=

b
m0

mn
i=1

b
mi
E[F
i
[ (27)
for some constants

b
mk
(k = 0; 1; 2; . . . ; mn) that for each m are functions of b
0
, b
1
, . . . , b
n
. Therefore, all required expectations
may be evaluated if the ow moments about the origin, up to the mnth power, may be evaluated. The same is also true of the
heteroscedastic expressions (19)(21), if we assume w() is also of polynomial form, since the product w(F)t(F) will also be
polynomial in F.
While writing the expansions of (27) to identify the constants is a little tedious, for special forms of the kind we shall
subsequently examine, their form is not so onerous. Specically, if we assume a powerlaw, BPR form of travel time function
(for which we simplify the notation a little, by setting b
0
= a and b
n
= b, with all other b
i
= 0 (i = 1, 2, . . . , n 1)):
t(f ) = a bf
n
(n a given positive integer) (28)
then:
E[t(F)[ = E[a bF
n
[ = a bE[F
n
[; (29)
E[(t(F))
2
[ = E[(a bF
n
)
2
[ = a
2
2abE[F
n
[ b
2
E[F
2n
[; (30)
E[(t(F))
3
[ = E[(a bF
n
)
3
[ = a
3
3a
2
bE[F
n
[ 3ab
2
E[F
2n
[ b
3
E[F
3n
[: (31)
Thus, drawing together the results of this section we have shown that in the case of a homoscedastic residual variance
and a power law, BPR travel time function, then the rst three travel time moments (mean, and second and third moments
about the mean) are given by (23)(25), and further that the expectations in these expressions may in turn be expressed in
terms of ow moments about the origin through (29)(31).
We have also shown that this method is extensible to more complex cases, with general nth order polynomial travel time
functions, whereby the travel time moments are given by (26) and (27), and heteroscedastic residual variances whereby
(19)(21) apply.
As mentioned at the start of this section, the reason for wishing to compute the theoretical travel time moments, and
relating them to ows, is that such expressions will form the basis of the method (described in Section 6) for correcting
778 D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789
observed travel times for growth in trafc ows. Therefore, with the expressions derived so far, we are part of the way to
achieving this goal. The remaining element to address is how to evaluate the require ow moments. Specically, focusing
on the homoscedastic/powerlaw case, how do we relate the computation of E[F
n
], E[F
2n
] and E[F
3n
] required in (29)(31)
to the model assumptions made? This is addressed in Section 5.
5. Computation of ow moments
The result
2
of Section 4 is that we can relate the travel time moments to the ow moments about the origin, E[F
n
], E[F
2n
] and
E[F
3n
]. In this section, we consider the computation of these ow moments, which is quite straightforward given the model
assumptions made, and uses several similar processes of expansion to those used in Section 4.
Now, dene the mth order ow moment about the origin as:
h
/
m
= E[F
m
[ (m = 1; 2; . . .): (32)
The moments in (32) can be calculated using a series of results. Firstly, we can relate these moments to yet another kind
of moment, namely the ow moments about the mean given by:
h
m
= E[(F E[F[)
m
[ (m = 2; 3; . . .): (33)
A binomial expansion inside the outer expectation in (33) yields:
(F E[F[)
m
=

m
j=0
(1)
mj
m!
j!(m j)!
F
j
(E[F[)
mj
(m = 2; 3; . . .) (34)
and then taking expectations of (34) to give (33), and using substitution (32), gives:
h
m
=

m
j=0
(1)
mj
m!
j!(m j)!
h
/
j
h
/
1
_ _
mj
(m = 2; 3; . . .); (35)
where we extend the denition of (32) naturally to the case m = 0, by assuming
h
/
0
= 1: (36)
A second result we now use is that for a variable F - Nor(l, r
2
), it can be shown that the mth order moment about the
mean is given by:
h
m
=
0 if m odd
m!
2
m=2
(m=2)!
r
m
if m even
_
(m = 2; 3; . . .): (37)
Taking these results together, since we can compute all moments about the mean directly from (37), then we can relate
them to the moments about the origin by (35). Examining (35), then for a given value of m, then h
m
is a linear combination of
h
/
1
; h
/
2
; . . . ; h
/
m
; looked at in reverse, if we already knew the values of h
/
1
; h
/
2
; . . . ; h
/
m1
(i.e. up to the (m 1)th moment about the
origin), then we could calculate the next moment about the origin h
/
m
explicitly from re-arranging (35) and using expression
(37) for h
m
. This suggests a recursive method for calculating the desired moments, whereby h
/
1
; h
/
2
; . . . are calculated in order.
From (29)(31) we know that the highest such moment we shall need is h
/
3n
= E[F
3n
[, where n is the power of the travel time
function (28), so that in fact we generate all moments up to this point, namely h
/
1
; h
/
2
; . . . ; h
/
3n
, and then at the end select those
that we need for substitution into (29)(31).
In particular, we re-express (35) slightly, by taking out the term corresponding to j = m:
h
m
= h
/
m

m1
j=0
(1)
mj
m!
j!(m j)!
h
/
j
h
/
1
_ _
mj
(m = 2; 3; . . .) (38)
and a slight re-arrangement of (38) yields:
h
/
m
= h
m


m1
j=0
(1)
mj
m!
j!(m j)!
h
/
j
h
/
1
_ _
mj
(m = 2; 3; . . .): (39)
The recursion suggested by (39) thus proceeds as follows:
1. Initialisation step. Set h
/
0
= 1 (by convention) and h
/
1
= l (the assumed mean). Initialise iteration counter m to 1.
2. Iteration counter. Increment m by 1.
3. Flow moment about mean. Calculate the mth moment about the mean h
m
from (37).
4. Flow moment about origin. Calculate the mth moment about the origin h
/
m
from (39).
5. Next step. If m < 3n, then return to step 2.
2
We restrict attention, for ease of illustration, to the case of a powerlaw travel time function and homoscedastic residual variance; the principles are just
the same for the case of polynomial travel time function and heteroscedastic residual variance, and still the method presented in this section applies.
D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789 779
For small values of n, we might derive explicit expressions from this recursion; for example:
h
/
2
= h
2
h
/
1
_ _
2
2 h
/
1
_ _
2
= r
2
h
/
1
_ _
2
= r
2
l
2
;
h
/
3
= h
3
l
3
3l
3
3h
/
2
l = 0 2l
3
3l(r
2
l
2
) = l
3
3lr
2
:
However, the real point of the recursion is to provide an implicit denition for large values of n, without having to derive
explicit expressions for all moments up to order 3n. By this method we are therefore implicitly expressing all such h
/
m
as
functions of l and r. Thus, the recursion becomes a very simple numerical procedure for successively generating all required
ow moments, which is applicable whatever value of n is used.
Having completed the recursion, we then select the appropriate ow moments for substitution into (29)(31), and in turn
substitute these expressions into (23)(25). This completes the process of expressing the travel time mean, variance and
third moment about the mean in terms of the assumed distribution of trafc ows. Comparing the travel time distributions
derived in the base/reference year and in the current year, we will then obtain a basis for correcting the currently-observed
travel times for trafc growth: this process is explained in the following section.
6. Method of correcting for trafc growth
Drawing together the results derived in Sections 4 and 5, in the present section we present a method that may be used to
make corrections to observed travel time data, to take out the effect of trafc growth that has occurred since some reference
year. Effectively, by analogy to re-basing of monetary values, we re-base the current year travel time observations to be in
comparable units to the base year observations, if trafc growth had not occurred.
Taking the key results of Sections 4 and 5 together, we have shown how the travel time moments /
/
1
; /
2
; /
3
_ _
may
through a combination of (23)(25), (29)(31) and the recursive method exploiting (37) and (39) be related to the travel
time function t(), the probability distribution of ows F, and the residual travel time variability. In terms of the parameters of
these latter entities, we have thus related /
/
1
; /
2
; /
3
_ _
to: the parameters of the travel time function (28), namely, a and b; the
parameters of the ow distribution l and r; and the residual variance as measured by b. Although we do not derive explicit
expressions relating these quantities, the combination of relationships and recursive method together dene an implicit rela-
tionship, and in order to reect this implicit dependence we may write (for some functions h
1
, h
2
and h
3
):
/
/
1
/
2
/
3
_
_
_
_
_
_ =
h
1
(l; r; a; b; b)
h
2
(l; r; a; b; b)
h
3
(l; r; a; b; b)
_
_
_
_
_
_: (40)
The implicit relationships in (40) provide the opportunity to estimate how much the travel time distribution (as mea-
sured by its moments /
/
1
; /
2
; /
3
) may be expected to change when any of the parameters l, r, a, b, b change.
This approach therefore has rather wide application. For example, it would be natural to reect capacity improvements
relative to some base year through a reduction to the parameter b, and therefore (40) provides the opportunity, say, to exam-
ine how the travel time distribution would be expected to change from a base year to a current year, if growth in trafc oc-
curs but the effect of the capacity improvement is taken out (i.e. how bad would conditions have become if the capacity
improvements had not taken place). We explore this interpretation further in Sections 7.2 and 7.3. There are clearly many
variants on this kind of interpretation.
However, our focus in the present section will be on separating out the effect of growth in trafc levels. Specically, we
shall write the mean ow level l relative to some ow level ~ l in the reference year:
l = k
~
l;
where k P1 may be used to reect growth in mean ow since the reference year. While the mean trafc level is our param-
eter of interest, we may also postulate whether (and if so, how) the other parameters in (40) might have changed between
the reference year and the current year, specically r, a, b and b. For example, it might be argued that an increase in mean
trafc levels would likely lead to an increase in the variability in trafc levels through r. Alternatively, there may be evidence
that the amount of residual travel time variation, as measured by b, has changed since the reference year. On the other hand,
there may have been policy measures that have affected the free-ow travel time a (such as enforcement policies on speed
limiters in heavy goods vehicles) or capacity-related parameter b. A natural rst approach followed here, however, is that
given the lack of strong evidence to the contrary, it is sensible to hold all other parameters except the mean trafc level con-
stant, and aim only to separate out the effect of this change from the observed travel time variability.
3
In this way, we may
use (40) to decompose each of the current year travel time moments into two components, namely the value of the moments in
the reference year (when l = ~ l in (40)), and the amount to which they are expected to change in the current year due to growth
in mean trafc ow (when l = k
~
l in (40)). Thus the travel time moments in the current year may be written:
3
If trafc growth really does lead to increased variance in ows, not just increased mean ow, then by assuming a constant variance we will systematically
underestimate the true impact of growth on the travel time distribution, so this can be claimed to be a rather conservative assumption.
780 D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789
/
/
1
/
2
/
3
_
_
_
_
_
_ =
h
1
(k
~
l; r; a; b; b)
h
2
(k
~
l; r; a; b; b)
h
3
(k
~
l; r; a; b; b)
_
_
_
_
_
_ =
h
1
(
~
l; r; a; b; b) (h
1
(k
~
l; r; a; b; b) h
1
(
~
l; r; a; b; b))
h
2
(
~
l; r; a; b; b) (h
2
(k
~
l; r; a; b; b) h
2
(
~
l; r; a; b; b))
h
3
(
~
l; r; a; b; b)
..
[Value in reference year
(h
3
(k
~
l; r; a; b; b) h
3
(
~
l; r; a; b; b))
..
Growth effect
_
_
_
_
_
_
_
_
_
_
: (41)
Suppose now that we observe sample estimates of the moments /
/
1
; /
2
; /
3
in the current year, and denote these estimates
by
^
/
/
1
;
^
/
2
;
^
/
3
. If we wish to correct these sample estimates for growth, then we need to subtract the growth effects in (41)
above. Thus we obtain corrected estimates, with the effect of growth taken out, which we denote
^
/
/CORR
1
;
^
/
CORR
2
;
^
/
CORR
3
and
which are given by:
^
/
/CORR
1
^
/
CORR
2
^
/
CORR
3
_
_
_
_
_
_
_
_
=
^
/
/
1
(h
1
(k
~
l; r; a; b; b) h
1
(
~
l; r; a; b; b))
^
/
2
(h
2
(k
~
l; r; a; b; b) h
2
(
~
l; r; a; b; b))
^
/
3
(h
3
(k
~
l; r; a; b; b) h
3
(
~
l; r; a; b; b))
_
_
_
_
_
_: (42)
To apply (42), we therefore go through the process of deriving the theoretical travel time moments twice, once for the
case where the mean trafc level is
~
l and once for the case where it is k
~
l, and this provides the implicit evaluation of
h
1
(~ l; r; a; b; b) and h
1
(k~ l; r; a; b; b), as required in (42). The extent to which these two quantities differ tell us the extent
to which growth since the reference year has affected the current observed travel times, and therefore provides the basis
for subtracting the impact of growth from the moments of the empirically observed distribution.
As a result, we end up with new estimates of the mean, variance and skewness that have the effect of growth taken out.
However, these summary measures are not exactly what is typically required of an evaluation such as a reliability assess-
ment, which would normally be based on percentiles of the travel time distribution, such as the 90th percentile travel time.
Therefore, the nal stage of the correction process will use these corrected moments to return to the original data, and cor-
rect the individual data points in the light of growth.
4
Thus, based on the corrected moment estimates, we begin by tting a parametric family of probability distributions to the
travel time moments. While one could t some quite exible family of curves, such as Johnson curves (see Clark and Watling,
2005) which may need an additional fourth moment, it should be adequate for our purposes to t something relatively easy
such as a lognormal. As described in this reference we can t a three-parameter lognormal of the form:
a d ln(X n) - Nor(0; 1) (43)
by the method of moments. The measures that we t to are the mean
^
/
/CORR
1
, the variance
^
/
CORR
2
, and a measure of skewness
derived from the third moment:
^
c
CORR
3
=
^
/
CORR
3
^
/
CORR
2
_ _
3=2
: (44)
To apply the method of moments, we rst solve for x the equation:
(x1)(x2) =
^
c
CORR
3
(45)
and if the solution is denoted by
^
x, we can then estimate the parameters of the lognormal distribution (43) from:
^
d = (ln
^
x
1=2
); (46)
^
a =
1
2
^
d ln
^
x(
^
x1)
^
/
CORR
2
_ _
; (47)
^
n =
^
/
CORR
1
exp
1
^
d
1
2
^
d

^
a
_ _ _ _
: (48)
From this tted distribution, the nal step is then to revise the observed intra-period mean journey times for each day,
through a normalisation method as follows:
1. Suppose that the uncorrected travel times for each of the N days are: {x
1
, x
2
, x
3
, . . . , x
N
}.
2. For these uncorrected travel times, calculate sample estimates of the three moments,
^
/
/
1
;
^
/
2
;
^
/
3
.
3. Fit a lognormal distribution to the uncorrected travel times by the method of moments, resulting in parameter estimates
of (d, a, n), denoted (
^
d
0
;
^
a
0
;
^
n
0
).
4. For the uncorrected travel times, form a sample of normalised uncorrected travel times {z
1
, z
2
, z
3
, . . . , z
N
}, where:
4
An alternative to the process described would be, after tting the parametric distribution, to compute the desired percentile theoretically, without resort to
the original data; however, we believe there to be an advantage in correcting the individual data points, since the analyst may then be in a better position to
judge the reasonableness of the correction applied, and in practice it will make hardly any difference to the nal computed percentile whichever method is
used.
D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789 781
z
i
=
^
a
0

^
d
0
ln(x
i

^
n
0
) (i = 1; 2; . . . ; N): (49)
5. Correct the sample moments for growth according to (42), yielding
^
/
/CORR
1
;
^
/
CORR
2
;
^
/
CORR
3
.
6. Fit a lognormal distribution to the corrected moments from step 5. by the method of moments, resulting in new param-
eter estimates of (d, a, n), denoted (
^
d;
^
a;
^
n) .
7. From the normalised uncorrected travel times, generate a sample of corresponding travel times corrected for growth,
denoted
~
x
1
;
~
x
2
;
~
x
3
; . . . ;
~
x
N
and given by:
~
x
i
=
^
n exp
z
i

^
a
^
d
_ _
(i = 1; 2; . . . ; N): (50)
Note that the normalisations correspond to the transformation of the assumed lognormal variable that gives a standard
Normal (0, 1) distribution. The process below is probably the most useful for implementing the method, though we could of
course combine steps 1, 4 and 7, by which we can see the relationship between each of the uncorrected data points x
i
and the
corresponding corrected data point ~x
i
, which after simplication can be expressed as:
~
x
i
=
^
n (x
i

^
n
0
)
^
d
0
=
^
d
exp
^
a
0

^
a
^
d
_ _
: (51)
However, the process is probably more transparent as a series of steps. Once the corrected travel times have been com-
puted, it is then straightforward to calculate, say, percentiles or any other measure from the corrected travel times, as if they
were the observed data. The procedure of adjusting the tted distribution is illustrated in Fig. 1.
7. Explicit formulae for second order travel time function
7.1. Decomposition of growth effects
The basis of the correction procedure (42) is made somewhat difcult to appreciate by the fact that, according to the pro-
cess described, the functions involved in (42) are only implicit, derived from the application of a recursive process and sev-
eral other steps from Sections 4 and 5. If, however, we are prepared to assume a second-order travel time function, then
some considerable simplication is possible. While such a function may not be a good t to real data, the purpose of this
section is more as a means of explaining the principles of the more general process, by deriving explicit formulae for this
special case.
Thus, in this section we assume a special case of a second order, powerlaw travel time function:
t(f ) = a bf
2
: (52)
Then simple, explicit formulae for the travel time moments of the mean /
/
1
, variance /
2
, and third moment about the
mean /
3
can be readily deduced. From (23)(25), we thus need to compute the travel time moments about the origin
(29)(31), which for n = 2 require knowledge of E[F
2
], E[F
4
], E[F
6
]. These ow moments can be shown to be:
P
r
o
b
a
b
i
l
i
t
y

d
e
n
s
i
t
y

Probability distribution of
uncorrected travel times
Probability distribution of
corrected travel times
90
th
percentile travel time
Corrections applied to observed
travel times
Travel time
Fig. 1. Adjustment of travel time distribution for growth.
782 D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789
E[F
2
[ = l
2
r
2
; (53)
E[F
4
[ = l
4
6l
2
r
2
3r
4
; (54)
E[F
6
[ = l
6
15l
4
r
2
45l
2
r
4
15r
6
: (55)
Substitution into (29)(31) yields:
E[t(F)[ = a b(l
2
r
2
); (56)
E[(t(F))
2
[ = a
2
2ab(l
2
r
2
) b
2
(l
4
6l
2
r
2
3r
4
); (57)
E[(t(F))
3
[ = a
3
3a
2
b(l
2
r
2
) 3ab
2
(l
4
6l
2
r
2
3r
4
) b
3
(l
6
15l
4
r
2
45l
2
r
4
15r
6
); (58)
and substitution into (23)(25) gives, after simplication:
/
/
1
= a b(l
2
r
2
); (59)
/
2
= b
2
2b
2
r
2
(2l
2
r
2
); (60)
/
3
= 8b
3
r
4
(3l
2
r
2
): (61)
Expressions (59)(61) are the explicit form of the system (40), meaning that we can combine all required expressions in
Sections 4 and 5 (including avoiding the need to implement the recursion given by (37)/(39)). In principle, we could antic-
ipate all values of the power n that might occur, and repeat this process for all such values, giving explicit formulae like (59)
(61) for each such value of n; however, the algebraic manipulation required would be enormous, and the implicit denition,
including the recursive denition of the ow moments, is then both algebraically and computationally attractive. Returning
attention to the special case of n = 2 we consider in this section, the relatively simplicity of the expressions (59)(61) is strik-
ing, making it possible to see many plausible, logical features. For example, we may see the way in which an increase in ow
variability r would increase all of the travel time mean, variance and third moment (and therefore the upper 90th percentile,
for example).
We now examine the impact of trafc growth; as in Section 6, we achieve this by assuming that
~
l denotes the mean traf-
c ow level in some base/reference year, and that for given k P1 the mean trafc ow in the current year is k~ l. Since (59)
(61) dene the functions in (40), we may apply (41) to give the decomposition of the moments into growth-related and
growth-independent parts, which after some simplication gives:
Mean travel time : /
/
1
= a b(
~
l
2
r
2
)
..
Value in reference year
b(k
2
1)
~
l
2
..
Growth effect
; (62)
Variance in travel time : /
2
= b
2
2b
2
r
2
(2
~
l
2
r
2
)
. .
Value in reference year
4b
2
r
2
(k
2
1)
~
l
2
..
Growth effect
; (63)
Skewness numerator : /
3
= 8b
3
r
4
(3
~
l
2
r
2
)
..
Value in reference year
24b
3
r
4
(k
2
1)
~
l
2
..
Growth effect
: (64)
We may then apply (42) to yield the method of correcting observed estimates of the moments for trafc growth, by sub-
tracting the growth effects, which just comes down to the simple equations:
^
/
/CORR
1
^
/
CORR
2
^
/
CORR
3
_
_
_
_
_
_
_
_
=
^
/
/
1
b(k
2
1)
~
l
2
^
/
2
4b
2
r
2
(k
2
1)
~
l
2
^
/
3
24b
3
r
4
(k
2
1)
~
l
2
_
_
_
_
_
_
_
_
: (65)
Given these corrected estimates, we may then apply the remainder of the procedure described in Section 6 to correct the
individual observations, and thus compute percentiles from the corrected observations as required.
7.2. Decomposition of capacity interventions in the light of trafc growth
Although the focus of this note has been on ultimately subtracting the effect of trafc growth from observed travel time
data, the methodology presented has wider applicability, as has been suggested on a number of occasions. To illustrate this
point, we consider again the second order form (52) in order to derive explicit expressions, even though the principle is quite
general. In particular, we will now aim for a different kind of decomposition. We suppose now that capacity improvements
have been made since the reference year, and that this may be reected by writing:
b = p
~
b; (66)
where
~
b is the value in the reference year of the parameter multiplying f
2
in (52), and p
~
b is the value of this parameter in the
current year, for some p 6 1. That is to say, an increase in capacity is naturally associated with a decrease in the b parameter
in (52).
D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789 783
Our aim will now be to examine the effect of trafc growth since the base year (as in Section 7.1) with and without the
capacity improvements; so that ultimately we can see what the travel time distribution would have looked like (in the face
of trafc growth), if the improvements actually made had not been made. The general approach is to modify (41) to examine
the effects with and without the capacity improvements:
/
/
1
/
2
/
3
_
_
_
_
_
_ =
h
1
(k
~
l; r; a; p
~
b; b)
h
2
(k
~
l; r; a; p
~
b; b)
h
3
(k
~
l; r; a; p
~
b; b)
_
_
_
_
_
_
_
_
=
h
1
(k
~
l; r; a;
~
b; b) (h
1
(k
~
l; r; a; p
~
b; b) h
1
(k
~
l; r; a;
~
b; b))
h
2
(k
~
l; r; a;
~
b; b) (h
2
(k
~
l; r; a; p
~
b; b) h
2
(k
~
l; r; a;
~
b; b))
h
3
(k
~
l; r; a;
~
b; b)
..
Value at reference year capacity
(h
3
(k
~
l; r; a; p
~
b; b) h
3
(k
~
l; r; a;
~
b; b))
..
Capacity improvement effect
_
_
_
_
_
_
_
_
_
_
_
_
: (67)
Note that in (67), all ows are evaluated at their current (growth) values, as we are now accepting that growth occurs but
factoring out the impact on the current year travel time distribution of the capacity improvements. Following the same ratio-
nale as for the case of factoring out trafc growth, it follows that in order to correct any currently-observed travel time mo-
ments for the effect of capacity improvements (i.e. how bad would they have been without the capacity improvements), then
we need to subtract the capacity improvement effects in (67) from the observed moment estimates, to get a capacity-cor-
rected set of moments (denoted CC):
^
/
/CC
1
^
/
CC
2
^
/
CC
3
_
_
_
_
_
_
_
_
=
^
/
/
1
(h
1
(k
~
l; r; a; p
~
b; b) h
1
(k
~
l; r; a;
~
b; b))
^
/
2
(h
2
(k
~
l; r; a; p
~
b; b) h
2
(k
~
l; r; a;
~
b; b))
^
/
3
(h
3
(k
~
l; r; a; p
~
b; b) h
3
(k
~
l; r; a;
~
b; b))
_
_
_
_
_
_
_
_
: (68)
Turning attention to the special case of second-order travel time functions (52), then we may again use expressions (59)
(61), to obtain in this case an explicit version of (67):
Mean travel time : /
/
1
= a
~
b((k
~
l)
2
r
2
)
..
Value at reference year capacity
(1 p)
~
b((k
~
l)
2
r
2
)
..
Capacity improvement effect
(69)
Variance in travel time : /
2
= b
2
2
~
b
2
r
2
(2(k
~
l)
2
r
2
)
..
Value at reference year capacity
2(1 p
2
)
~
b
2
r
2
(2(k
~
l)
2
r
2
)
..
Capacity improvement effect
(70)
Skewness numerator : /
3
= 8
~
b
3
r
4
(3(k
~
l)
2
r
2
)
..
Value at reference year capacity
8(1 p
3
)
~
b
3
r
4
(3(k
~
l)
2
r
2
)
..
Capacity improvement effect
: (71)
Therefore for a second order travel time function, the correction in (68) takes the form:
^
/
/CC
1
^
/
CC
2
^
/
CC
3
_
_
_
_
_
_
_
_
=
^
/
/
1
(1 p)
~
b((k
~
l)
2
r
2
)
^
/
2
2(1 p
2
)
~
b
2
r
2
(2(k
~
l)
2
r
2
)
^
/
3
8(1 p
3
)
~
b
3
r
4
(3(k
~
l)
2
r
2
)
_
_
_
_
_
_
_
_
: (72)
Note that in this case (in contrast to Section 7.1) the correction adds something positive to all moments rather than sub-
tracts, since the correction is aiming to show how much worse things would have become in the current year, without the
capacity improvements.
7.3. Decomposition of variance components: capacity interventions and trafc growth
While we have dealt separately with trafc growth and capacity improvements in Sections 7.1 and 7.2, it is only natural to
ask whether we may deal with them simultaneously in any sensible way. The answer is yes. While we could aim to make a
correction to the observed data to show what would have happened had growth not occurred and had the capacity improve-
ments not occurred (and such a correction naturally follows from the decomposition below), it is probably more instructive
to nish at the step before the correction is made, by examining the decomposition of the moments in the light of these
factors.
The approach for making this decomposition follows a natural extension of the ideas already presented in Sections 7.1
and 7.2, and here we simply present and comment on the explicit results (see Table 1) that can be readily derived from
(59)(61) for the case of a second order travel time function:
Thus, depending on which correction terms are applied to the observed travel time moments, we may choose to correct
(a) for growth only, (b) for capacity improvements only, or (c) for both growth and capacity improvements. Note, however,
that these effects are not purely additive due to the last column in the table above that involves both growth and capacity
improvement effects. That is to say, it is not correct to rst adjust the moments for growth, and then subsequently to adjust
those corrected moments for capacity improvements, since then we will effectively double-count the fourth term. Correcting
for both growth and capacity improvements together is best done by returning to the original, uncorrected observed mo-
ments, and then subtracting all three of the correction terms in columns 24 of the table above.
784 D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789
8. Numerical application of the model to assessing vulnerability of travel times to network growth
This section illustrates the principles described in Section 6 of this paper. As described in Section 1, the numerical illus-
trations take a retrospective view initially to obtain the model parameters and then assess the vulnerability of road links for
growth in demand in the future. In order to illustrate the method, we now need to consider specic types of roads with
known speed ow relationships. In transport modelling, commonly, we use standard speedow relationships which can
be easily converted to travel timeow relationships, provided the lengths of road sections are known. The procedure for
converting standard speedow curves from the NTM FORGE (DfT, 2005) is described in the next section.
8.1. Fitting ow-delay curves to NTMs speedow curves
In this research we have tted ow-delay curves to standard speedow relationships for UK roads which are used in
various transport modelling applications, and notably the UK National Transport Model (NTM; DfT, 2005). This approach
facilitates us illustrating the travel time adjustment method in application to various common types of roads found in prac-
tical planning studies (e.g., A-roads, single, dual carriageways). Table 2 shows the trafc ows and speeds used to derive
speedow curves for motorways by the NTM. A complete set of values for different types of roads in England can be found
in Table 9 of the NTM Report (DfT, 2005) but in the following table, we show the value for a Motorway.
The standard NTM relationship is derived by assuming a linearly decreasing speed between the corresponding values
of ow indicated in Table 2. In order to convert this speedow relationship into a travel timeow relationship, we
need to identify the characteristics of each link, e.g. its length. The travel time relationship to calibrate is an nth order
power law of ow:
t(f ) = a b f
n
;
Table 2
Motorway speedow values used to derive standard NTM
relation.
Speed (kmph) Flow (PCU/hour/lane)
115.8 0
112.6 1398
55.6 2330
20 2913
Source: Table 9, NTM FORGE model (DfT, 2005).
Table 1
Impact of capacity improvements and trafc growth.
D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789 785
where t(f) is the travel time in seconds at a ow f, a is the free ow time in seconds, f is the ow in PCUs/hour/lane, and n is a
positive integer. The constant b is calibrated from the knowledge of the capacity (in PCUs/hour/lane) and the capacity time
(travel time at capacity, in seconds) by rearranging the travel time expression at a ow = capacity:
b =
Capacity Time a
Capacity
n
:
From Table 2 and for a given stretch of a Motorway, we are able to deduce the free ow time, i.e. the travel time at no (or
low) ow, and also the capacity expressed in PCUs/hour/lane and the corresponding time at that capacity. Therefore, we
know all other information to t a travel time function except the parameter n; for any given value of n, we can calculate
the appropriate values of a and b from the data given. The selection of an appropriate value of n was achieved by a trial
and error method. A typical travel timeow relationship tted to the NTM speeddelay curve is shown in Fig. 2.
8.2. Numerical results
In this section, we consider some typical road stretches purely for illustration purposes, however, they do not necessarily
represent any existing roads. Numerical illustrations given below assume road stretches of specic types with known char-
acteristics such as road length, summaries of ow (viz., mean ow, ~ l in PCUs/hour/lane and its variance, r
2
) and travel time
(viz., mean travel time
^
/
/
1
sec, its variance,
^
/
2
and skewness,
^
/
3
). Each numerical example set out below considers various
scenarios of growth in trafc demand and assesses the vulnerability of a road stretch for demand growth. For the purpose
of this illustration, we dene the Travel Time Reliability (TTR) as a percentage change in mean, Standard Deviation (SD) and
skewness, relative to the base. More formally, TTR measures can be dened as below using the notation previously
introduced.
% Change in Mean Travel Time; W
1
=
^
/
/CORR
1

^
/
/
1
_ _
100=
^
/
/
1
;
% Change in SD of Travel Time; W
2
=
^
/
/CORR
2

^
/
/
2
_ _
100=
^
/
/
2
;
% Change in Skewness of Travel Time; W
3
=
^
/
/CORR
3

^
/
/
3
_ _
100=
^
/
/
3
:
Let the critical values of TTR for mean, SD and skewness be 1%, 10% and 1% respectively, which means that in the future if
the TTR exceeds any of the critical values, then the road stretch is said to be vulnerable for growth in demand. Finally, let us
dene the term, Demand Growth Reliability Vulnerability (DGRV) as the maximum growth rate j% pa which can result in
TTR being greater than the critical values identied earlier. For the purpose of illustration, we assume a horizon period of
three years, denoted by j(=3 in all cases, unless stated otherwise) over which the vulnerability of a road stretch is being as-
sessed. In the methodology section, we have dened a parameter k to indicate the growth in demand which can be related to
growth rate j% pa by using a simple relation, k=(1+j/100)
j
.
0
200
400
600
800
1000
1200
1400
1600
0 500 1000 1500 2000 2500 3000 3500
T
r
a
v
e
l

t
i
m
e

i
n

S
e
c
Flow in PCU/hr/lane
Travel Time- Flow Relationship for A Typical Motorway Link
t = 234 + (1.83416 10
-18
) (flow)
6
NTM Curve
Fig. 2. Fitted travel timeow relationship.
786 D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789
8.2.1. Numerical Example 1: (B-road)
Firstly, let us consider a busy B-type road
5
of length 3 km (1.864 miles) connecting two A-type roads. UK Department for
Transports NTM FORGE Model (DfT, 2005) species the speedow relationships for a hierarchy of roads including B-type roads
which can be readily converted into a travel timeow relationship of type (28) for the road stretch being considered. The
parameters of the relationship are given below. Note that the value of b depends on the units in which the ow has been mea-
sured and hence due care should be taken to ensure their consistency with each other.
a = 140 s;
~
b = 9:81023 10
5
(up to5 d:p:);
n = 2:
Let us assume that the mean ow on this road in the current year,
~
l =1000 PCUs/hour/lane with a variance, r
2
= 8000. Let
us also assume that the observed mean travel time
^
/
/
1
= 200 s with a variance
^
/
2
= 12; 000 s
2
and skewness
^
/
3
= 10. Con-
sidering various growth scenarios over the horizon period of 3 years, TTR measures have been computed and are summa-
rised in Table 3. TTR measures were then compared to the corresponding critical values and vulnerability of the road
stretch is readily assessed for each growth scenario. The mean travel time marginally increases with growth in demand,
but the SD increases signicantly over the base situation. The skewness reduces slightly with increase in demand reecting
the signicant increase in SD. When the trafc grows at j = 3% pa, the critical TTR value is exceeded and hence the road
stretch is considered to be vulnerable to trafc growth at that level. A simple maximisation of j yields the vulnerability
of the road stretch to be DGRV = 2.487 which means that the road is vulnerable to percentage demand growth rates above
this value.
8.2.2. Numerical Example 2: (Motorway stretch)
A second numerical example has been constructed in much the same way as before, but this time for a high capacity
motorway stretch of 44 km in length. Using DfTs NTM information, a travel timeow relationship has been derived and
the parameters for this stretch of motorway are as indicated below.
a = 170 s;
~
b = 1:39537 10
18
(up to 5 d:p:);
n = 6:
Let us assume that the rush hour ow and travel time summaries in the base year are known which are given below:
~
l = 1800 PCUs=hour=lane;
r
2
= 5000;
^
/
/
1
= 200 s
^
/
2
= 15; 000 s
2
^
/
3
= 8:5:
5
B-roads in England are usually regional in nature connecting non-primary A-type roads. Essentially, they are low capacity roads with undivided
carriageways with at-grade junctions which may be signal controlled. A-type roads are relatively higher capacity roads and primary A-roads may even have
dual carriageways and grade separations.
Table 3
Vulnerability analysis of a B-road stretch.
TTR measure Growth rate of demand (j% pa)
1% 2% 3%
W
1
0.00144 0.00283 0.00416
W
2
4.52259 8.34861 11.61321
W
3
0.02664 0.05453 0.08372
Assessment Not Vulnerable Not Vulnerable Vulnerable
Table 4
Vulnerability analysis of a motorway stretch.
TTR measure Growth rate of demand (j% pa)
0.1% 0.25% 0.5%
W
1
0.42967 1.05915 2.06952
W
2
0.01282 0.03133 0.06034
W
3
0.02134 0.05344 0.10719
Assessment Not Vulnerable Vulnerable Vulnerable
D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789 787
Note that the trafc ow is higher than in the previous example, and the variance in travel time is also signicantly higher
reecting a large spread of travel times on this busy motorway stretch. A slightly lower, but positive skewness indicates that
some drivers have experienced extremely long travel times on this road. TTR measures were then computed using identical
procedures as before considering various growth scenarios. However, the critical growth rate is much lower compared to the
previous example as the motorway stretch under consideration cannot simply accommodate such a high degree of growth.
Table 4 shows the assessment of vulnerability. Again, by maximising j subject to the condition that the TTR measure is less
than the critical TTR value yielded DGRV = 0.2357. This stretch of motorway is therefore vulnerable to growth in demand
above 0.2357%.
In a situation like this, perhaps, a small increase in road capacity, rather than an addition of a complete lane, may be sig-
nicant in reducing vulnerability to growth. Modern management of highway operations involves increasing the throughput
by adopting some measures such as deploying trafc ofcers on motorways to dampen the unstable ow conditions. Such
measures might be assimilated to small year-on-year increases in capacity of the roads. Let us denote the % increase in capac-
ity per year by q which can be related to the parameter p dened earlier by p = [(1 + q)
j
1]. A 1% annual increase in motor-
way capacity, i.e. q = 1% (together with assuming 0.5% annual increase in demand) would ease the situation by reducing the
change in mean travel time W
1
to 1.39647%. Corresponding changes in SD and skewness worked out as, W
2
= 0.03709 and
W
3
= 0.10718. A further increase in capacity to 2% would make the road stretch less vulnerable to growth in demand by
bringing W
1
to below 1% which was the assumed critical TTR value. In this latter case with capacity expansion, DGRV works
out as 0.5696 above which the road is vulnerable to growth in demand. This analysis clearly shows that the interventions
made by the authorities could help counter the ination in trafc ows.
9. Concluding remarks
In this paper, we have argued that growth in demand is a relevant factor to consider when assessing the vulnerability of a
road network, since it will tend to mean the network will operate in a regime closer to capacity, where small variations in
ow or capacity may have major impacts on travel times (and thus on the reliability of travel times). Based on this obser-
vation, a quite simple methodology has been proposed, which effectively tracks moments of the ow and travel time distri-
bution, as a means of separating out the effect of trafc growth on link travel time moments. Such a method can be applied to
historic data, to identify the contribution of past growth in demand on observed travel time moments. However, the focus of
the present paper has been to utilise this technique in a forward-looking manner, whereby the planner species critical
changes in the travel time moments, and the model is used to identify the vulnerability of links in terms of the demand
growth they can accommodate before breaching these critical levels. The resulting measure, termed Demand Growth Reli-
ability Vulnerability, may be attributed to each link of a transport network in order to identify levels of ow growth to which
each link is vulnerable. An example application of this method has been given to two prototypical road links.
The method presented may be extended in a number of ways in the future. As noted earlier, the link-level analysis of vul-
nerability may be integrated into a network-level analysis, by assessing now the vulnerability to OD demand growth, follow-
ing an adaptation of the methods in Clark and Watling (2005) developed for the case of Travel Time Reliability to apply to
the vulnerability analysis. The method may further be developed by assuming within-day dynamic link travel time func-
tions, for which a Taylor series approximation may be used to derive polynomial relationships between travel time and ow,
utilising the calculus techniques developed in Balijepalli and Watling (2005). Finally, we may bring together the analysis of
capacity and demand growth vulnerability into one combined framework, whereby both sources are simultaneously
analysed.
Acknowledgements
The authors gratefully acknowledge the support of the UK Highways Agency for sponsoring the research that laid the
groundwork for the methods reported in this paper. However, the views expressed in the paper are those of the authors
alone, and should not be taken as reecting any view of the sponsor. The authors thank Dr. Tom Van Vuren for his comments
and suggestions on the modelling approach, and the comments of three anonymous referees which helped to improve the
presentation of the paper.
References
Balijepalli, C., Watling, D., 2005. Doubly dynamic equilibrium distribution approximation model for dynamic trafc assignment. In: Mahmassani, H. (Ed.),
Transportation and Trafc Theory: Flow, Dynamics and Human Interaction. Elsevier, Oxford, UK, pp. 741760.
Bell, M.G.H., 2000. A game theory approach to measuring the performance reliability of transport networks. Transportation Research Part B 34 (6), 533546.
Bell, M.G.H., Cassir, C., 2000. Reliability of Transport Networks. Research Studies Press, Baldock, Herts.
Bell, M.G.H., Iida, Y., 1997. Transportation Network Analysis. Wiley, Chichester.
Berdica, K., 2002. An introduction to road vulnerability: what has been done, is done and should be done. Transport Policy 9, 117127.
Chen, A., Yang, H., Lo, H.K., Tang, W.H., 2002. Capacity reliability of a road network: an assessment methodology and numerical results. Transportation
Research Part B 36 (3), 225252.
Clark, S., Watling, D., 2005. Modelling network travel time reliability under stochastic demand. Transportation Research Part B 39 (2), 119140.
DEste, G.M., Taylor, M.A.P., 2001. Modelling network vulnerability at the level of national strategic transport network. Journal of East Asia Society for
Transport Studies 4 (2), 114.
788 D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789
DfT, 2005. <http://www2.dft.gov.uk/pgr/economics/ntm/> (accessed 02.07.11).
DfT, 2007. Congestion on the strategic road network. 2004/5 PSA target base line gure and methodology.
FHWA, 2005. <http://ops.fhwa.dot.gov/publications/tt_reliability/brochure/> (accessed on 03.07.11).
Jenelius, E., 2009. Network structure and travel patterns: explaining the geographical disparities of road network vulnerability. Journal of Transport
Geography 17, 234244.
Jenelius, E., Petersen, T., Mattsson, L.-G., 2006. Importance and exposure in road network vulnerability analysis. Transportation Research Part A 40, 537560.
Nicholson, A.J., Dalziell, E., 2003. Risk evaluation and management: a road network reliability study. In: Bell, Iida (Ed.), The Network Reliability of Transport.
Pergamon-Elsevier, Oxford, pp. 4559.
Rijkswaterstaat, 2005. <http://www.rijkswaterstaat.nl/images/Travelling%20Time%20Reliability_tcm174-275339.pdf> (accessed 03.07.11).
Shao, H., Lam, W.H.K., Tam, M.L., 2006. A reliability-based stochastic trafc assignment model for network with multiple user classes under uncertainty in
demand. Networks & Spatial Economics 6, 173204.
Sumalee, A., Kurauchi, F., 2006. Reliability and emergency issues in transportation network analysis. Networks & Spatial Economics 6 (34), 169172.
Sumalee, A., Watling, D.P., 2008. Partition-based algorithm for estimating transportation network reliability with dependent link failures. Journal of
Advanced Transportation 42 (3), 213238.
Szeto, W.Y., OBrien, L., OMahony, M., 2006. Risk-averse trafc assignment with elastic demands: NCP formulation and solution method for assessing
performance reliability. Networks & Spatial Economics 6, 313332.
Taylor, M.A.P., Somenahalli, V.C.S., DEste, G.M., 2006. Application of accessibility based methods for vulnerability analysis of strategic road networks.
Networks & Spatial Economics 6, 267291.
Yang, H., Lo, H.K., Tang, W.G., 2000. Travel time versus capacity reliability of a road network. In: Bell, M.G.H., Cassir, C. (Eds.), Reliability of Transport
Networks. Research Studies, Baldock, Herts, pp. 119138.
D. Watling, N.C. Balijepalli / Transportation Research Part A 46 (2012) 772789 789